BROWSE
Detail
Conference Paper
Egocentric 3D Pose Estimation with Temporally Guided Pose Refinement
Cited - time in 
Share 
Share
- Authors
- Seongmin Baek, Youn-Hee Gil, Hyunjun Lim
- Issue Date
- 2025-12
- Citation
- ACM SIGGRAPH Asia (SA) 2025, pp.1-3
- Publisher
- ACM
- Language
- English
- Type
- Conference Paper
- Abstract
- Unlike single-frame pose estimation, human motion is inherently temporally continuous, and this challenge is further compounded by egocentric views, which suffer from severe self-occlusions and limited FoV, making per-frame inference unstable. We address these issues with a two-stage pipeline: a causal Temporal Convolutional Network (TCN) over the previous T frames predicts a motion-conditioned pose prior, and a single-layer Transformer refines this prior with current egocentric visual features to produce the final 3D pose. This design yields accurate and temporally coherent egocentric pose estimation with minimal latency.
- KSP Keywords
- 3D pose estimation, Convolutional networks, Human motion, Single-layer, Two-Stage, Visual Features